According to a Sankey diagram of model usage among T3 Code users shared by Jeremy Howard on X, within three days of Gemini 4 Neon's launch (Nov 18–20), Neon handled more than half the conversation turns in 6,330 of 16,341 installs. The data shows Neon drew share mainly from Gemini's own users and Opus users, and is being treated as empirical evidence of a shifting model landscape.

In the same post, Jeremy Howard also relayed a comment noting that when Neon launched, apparently no one inside Google DeepMind realized how strong it was, and the company repeatedly described it only as "close to" frontier models.

User migration data: Opus and Gemini's own users are the main sources

The Sankey diagram further breaks down conversion by prior model. Of the 5,306 installs that had used Opus 5.5 most, 9% shifted to Neon; users of other Gemini models had the highest conversion rate at 63%; GPT-61 Sol users came in at 10%, and users of other Claude models at just 8%.

By flow, Neon's gains are concentrated in two groups: existing users within the Gemini ecosystem, and Claude users who had been using Opus 5.5. That distribution corroborates the editorial judgment that "Neon took a large share of Opus users within three days of launch."

Industry reaction: official framing seen as an underestimate, benchmarks questioned

In his comment, Jeremy Howard cited Perplexity CEO Arav Srinivas, who said Neon, with reasoning set to High, is a rare model that is both "smart and fast," and that Srinivas has already moved some team members onto it.

Jeremy concluded from this that benchmarks are a very imperfect tool for measuring a model's usefulness.

That view echoes the actual migration data: if the official description frames Neon merely as "close to" frontier models while real-world usage shows large-scale switching, then existing benchmarks may fail to fully capture the model's practical value.

What to watch next

  • Whether Neon's user share keeps rising after Nov 20, and whether the high share across 6,330 installs holds steady.
  • Whether the churn rate among Opus 5.5 and other Claude model users widens further.
  • Whether Google DeepMind adjusts its official positioning and benchmarking of Neon.
  • Whether more third-party usage data can validate the judgment that "benchmarks are distorted."